Rui Wang

Would be possible to unify the package name and Cuda-version have in a version like PT, for example, torch==1.11.0+cu113 instead of bagua-cuda113==0.9.0? https://github.com/PyTorchLightning/pytorch-lightning/pull/12723

support amp with fused optimizer

2

torch native amp + apex amp

qadam algorithm occasionally failed in CI using `baguasys/bagua:master-pytorch-1.9.1-cuda11.1-cudnn8` image

int32 ...

We can add a dataloader/dataset wrapper, which caches item -> preprocessed results, to accelerate data loading. - [x] redis backend - [ ] rocksdb backend

The code is copied partially from #656

fix Lightning failing Bagua test for manual optimization

1

see this：https://github.com/Lightning-AI/lightning/pull/16225

[Draft] ci: fix bagua ci problems

1

Open this pr to fix bagua ci problems